diff --git a/pandagg/response.py b/pandagg/response.py
index 62075407..94507e04 100644
--- a/pandagg/response.py
+++ b/pandagg/response.py
@@ -155,9 +155,9 @@ def get(self, key):
     def _parse_group_by(
         self,
         response,
+        until,
         row=None,
         agg_name=None,
-        until=None,
         ancestors=None,
         row_as_tuple=False,
         with_single_bucket_groups=False,
@@ -166,11 +166,14 @@ def _parse_group_by(
 
         Yields each row for which last bucket aggregation generated buckets.
         """
+        # initialization: find ancestors once for faster computation
         if ancestors is None:
             ancestors = self._aggs.ancestors(until, id_only=True)
+            # remove eventual fake root
+            ancestors = [until] + [a for a in ancestors if a != "_"]
+            agg_name = ancestors[-1]
         if not row:
             row = [] if row_as_tuple else {}
-        agg_name = self._aggs.root if agg_name is None else agg_name
         if agg_name in response:
             agg_node = self._aggs.get(agg_name)
             for key, raw_bucket in agg_node.extract_buckets(response[agg_name]):
diff --git a/tests/test_response.py b/tests/test_response.py
index 2b504779..3228feb9 100644
--- a/tests/test_response.py
+++ b/tests/test_response.py
@@ -271,6 +271,16 @@ def test_parse_as_tabular_multiple_roots(self):
             },
         )
 
+        # with specified grouped_by
+        index_names, index_values = Aggregations(
+            data=raw_response, search=Search().aggs(my_agg)
+        ).to_tabular(grouped_by="classification_type")
+        self.assertEqual(index_names, ["classification_type"])
+        self.assertEqual(
+            index_values,
+            {("multiclass",): {"doc_count": 439}, ("multilabel",): {"doc_count": 433}},
+        )
+
     def test_parse_as_dataframe(self):
         my_agg = Aggs(sample.EXPECTED_AGG_QUERY, mapping=MAPPING)
         df = Aggregations(